An intelligent wheelchair bed posture regulation method and system, medium

Abstract

The application discloses a kind of intelligent wheelchair bed posture regulation method and system, medium, it is related to intelligent wheelchair bed technical field, comprising the following steps: construct discrete state space equation for MPC prediction;Multi-objective MPC cost function is constructed, combined with discrete state space equation, based on the optimal control increment sequence that can minimize cost function is solved out by the rolling horizon of sequence quadratic programming;The difference between the actual execution angle increment that is determined by optimal control increment sequence and the real-time feedback of bottom layer motor encoder in cycle is used as control deviation, input to corresponding PID controller to realize posture regulation.The application adopts the double-layer closed-loop architecture of MPC model prediction control combined with high-frequency PID, and the high-precision strong robust posture regulation of actuator is realized by the combination of double loops, the comfort quantitative evaluation item is written into MPC cost function, so that the comfort is changed from subjective fuzzy evaluation to objective target that can be calculated and optimized, actively avoids stress concentration when trajectory tracking, and guarantees the safety and comfort of riding.

Description

Technical Field

[0001] This invention relates to the technical field of intelligent wheelchair beds, and more particularly to an intelligent wheelchair bed posture control method, system, and medium. Background Technology

[0002] Virtual nursing homes are a new form of elderly care that has emerged in response to the trend of home-based elderly care. Compared to traditional nursing homes with physical walls built of bricks and cement, virtual nursing homes rely on digital technology to create "digital walls," building a safety protection network and service dispatch system covering home-based elderly care.

[0003] Among the elderly population using virtual nursing homes for home-based care, the care of disabled and semi-disabled seniors presents significant challenges. In addition to requiring professional caregivers to provide in-home services, equipping this group with intelligent equipment offering multi-scenario care functions—such as relocation, transfer, mobility, and safety monitoring—is crucial for enhancing their independence and autonomy. To address this, researchers have developed an intelligent wheelchair bed system integrating a nursing bed, wheelchair, relocation function, and vital sign monitoring. This system allows seniors to transfer and move without being lifted, effectively reducing the risk of secondary injury during care and enhancing their autonomy and sense of dignity.

[0004] Current wheelchair beds mostly rely on preset trajectories or simple logic control for posture changes, lacking real-time perception of changes in human sitting posture and body pressure distribution, making it difficult to achieve adaptive adjustment in human-machine collaboration. Especially during bed-chair transitions, and lacking prediction and optimization of the posture at the next moment, problems such as mechanical shock and stiff movements often occur, causing transient discomfort and even safety hazards for the elderly. Summary of the Invention

[0005] The purpose of this section is to outline some aspects of embodiments of the present invention and to briefly describe some preferred embodiments. Simplifications or omissions may be made in this section, as well as in the abstract and title of this application, to avoid obscuring the purpose of these documents; however, such simplifications or omissions should not be construed as limiting the scope of the invention.

[0006] To address the technical problem that existing intelligent wheelchair beds struggle to achieve adaptive adjustment and mitigate transient discomfort through human-machine collaboration, this invention provides a posture control method for an intelligent wheelchair bed. The invention offers the following technical solution, comprising the following steps: Constructing discrete state-space equations for MPC prediction: ; in, for State vector at time: ; To control the input vector: ; In the formula, These represent the actual feedback angles of the lower leg plate, upper thigh plate, and back plate at time k. This corresponds to the angular velocity; These represent the expected angle increments sent to the lower leg plate, upper leg plate, and back plate motors within control cycle k, respectively. Constructing the multi-objective MPC cost function J: ; In the formula, The actual angle vectors of the lower leg plate, leg plate, and back plate; This is the corresponding reference planning angle vector; To solve for the control increment sequence, The overall objective comfort score is estimated based on the current and future predicted postures. Q and R are the corresponding weight matrices, w is the comfort deviation penalty coefficient, and N is the prediction step size. Combining discrete state-space equations, the cost function is solved in the rolling time domain based on sequential quadratic programming. The minimized optimal control increment sequence; Each of the backrest motor, thigh board motor, and calf board motor is equipped with an independent standard PID controller. The difference between the actual execution angle increment determined by the optimal control increment sequence and the real-time feedback of the underlying motor encoder within the cycle is used as the control deviation and input to the corresponding PID controller to achieve posture control.

[0007] As a preferred embodiment of the intelligent wheelchair bed posture control method described in this invention, wherein: The acquisition method is as follows: the actual angle of each plate is fed back in real time by angle sensors deployed in the joints; key pressure feature parameters including peak body pressure, average pressure, contact area and pressure gradient are extracted by pressure sensors embedded in the seat cushion and backrest; the feature parameters are input into the comfort evaluation model based on random forest to output the current moment's comfort quantification score in real time.

[0008] As a preferred embodiment of the intelligent wheelchair bed posture control method described in this invention, the method further includes: constructing a CNN-LSTM autoencoder network based on spatiotemporal feature fusion, and using this CNN-LSTM autoencoder network to predict the human body contact pressure distribution map at the next moment; specifically including: A convolutional neural network was used as the encoder, and spatial topological features were extracted independently from the stress maps of five consecutive frames through temporally distributed operations. The input matrix was sequentially processed through two steps with a stride of 2 and padding of 1. Convolutional layer; the first convolutional layer expands the number of channels to 16, and the feature map size is downsampled to [value missing]. The second convolutional layer further reduces its dimensionality to... The feature map is flattened into a one-dimensional vector of length 2^112, and then compressed to a low-dimensional feature representation of 60 dimensions by a fully connected layer. The extracted spatial feature vector sequences at 5 independent time points are sequentially input into the Long Short-Term Memory network to output the final hidden state vector h(t). The decoder is responsible for accurately mapping the low-dimensional spatiotemporal representation h(t) back to the original physical space dimension, generating the predicted image: the 60-dimensional features are upscaled to 2^112 through a fully connected layer and reshaped into... A high-dimensional tensor; upsampling and reconstruction are performed through two transposed convolutional layers; the first transposed convolutional layer outputs... The feature map is then precisely restored to a single-channel predicted pressure matrix P(t+1) consistent with the dimensions of the real sensor in the second layer. After the network generates a high-resolution predicted stress matrix P(t+1), key stress features are extracted from this predicted matrix. Then, a random forest evaluation model is used to calculate the overall comfort prediction score for future postures. .

[0009] As a preferred embodiment of the intelligent wheelchair bed posture control method described in this invention, the method combines discrete state-space equations and solves the cost function in the rolling time domain based on sequential quadratic programming. When minimizing the optimal control increment sequence, a linear inequality constraint is set for the control increment command at each step in the prediction time domain: ; in, This indicates the control amount for a single adjustment of the backplate. This indicates the control amount for a single adjustment of the legboard. This indicates the control amount of a single adjustment of the calf plate. , representing the control quantity at each step within the finite prediction time domain in the future.

[0010] In a preferred embodiment of the intelligent wheelchair bed posture control method described in this invention, the control law of the PID controller is designed as follows: ; In the formula, e(t) represents the control deviation. For controller output, To control deviation, , , These are all parameters for a PID controller.

[0011] As a preferred embodiment of the intelligent wheelchair bed posture control method described in this invention, the method further includes: Based on the target posture parameters input by the user, determine the target tilt angle of the backrest relative to the seat surface. And the target folding angle of the thigh and calf plates relative to the seat surface. and ; The entire adjustable angle space of the wheelchair is discretized into a grid according to the set step length, and the continuous two-dimensional angle space is divided into a finite number of discrete grid points. The discomfort cost of all grid points is predicted using the SVR surrogate model, and an discomfort potential field covering the global space is generated. Each discrete grid point is regarded as a vertex of a weighted graph. The motion direction between vertices is defined by an 8-neighborhood connection method. The discomfort score of adjacent vertices is used as the weight of the edge. The network graph is established in combination with the physical constraints of the motor's single-step operation. Dijkstra's algorithm is used to perform a global search in the network graph, and the resulting discrete path node sequence is planned. It will automatically avoid high-load areas in the potential field; The planner combines the motor's maximum angular velocity and angular acceleration constraints, and uses a cubic spline interpolation algorithm to fit the discrete optimal path to generate a reference trajectory sequence: This is used as the penalty benchmark for the first term in the MPC cost function.

[0012] As a preferred embodiment of the intelligent wheelchair bed posture control method of the present invention, the method further includes: constructing a motor transfer function model for PID control, specifically: By applying drive signals with different PWM duty cycles to each channel motor, the angle change data of the motor is collected at a high frequency cycle. Based on the collected angle change data, the first-order or second-order transfer function models of the backrest, thigh board, and calf board motors are identified. Based on the obtained transfer function models, the Ziegler-Nichols method is used to calculate the exclusive PID parameter combinations of each motor under the two working conditions of lifting and lowering.

[0013] As a preferred embodiment of the intelligent wheelchair bed posture control method of the present invention, the weight matrix Q, R and the penalty coefficient w are globally adaptively optimized using a Bayesian optimization method.

[0014] This invention also discloses an intelligent wheelchair bed posture control system, applied to the aforementioned intelligent wheelchair bed posture control method, the system comprising: The planner module combines the maximum angular velocity and angular acceleration constraints of the motor, uses a cubic spline interpolation algorithm to fit the discrete optimal path to generate a continuous and smooth time reference trajectory sequence, and uses it as the penalty benchmark for the first term in the MPC cost function. The MPC controller module is used to combine discrete state-space equations with the MPC cost function to solve for the optimal control sequence output in the rolling time domain based on sequential quadratic programming. The spatiotemporal feature fusion network module is used to generate a high-resolution predicted stress matrix. Based on the predicted stress matrix, stress features are extracted and fed into the trained random forest evaluation model to calculate the overall comfort prediction score under future postures. This score is used as the active comfort optimization cost term in the MPC cost function. The system also includes a PID controller module, which uses the difference between the expected control quantity of the optimal control sequence and the actual execution angle increment fed back in real time by the underlying motor encoder within the cycle as the control error input to the PID control law to perform wheelchair bed posture adjustment.

[0015] The present invention also discloses a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described intelligent wheelchair bed posture control method.

[0016] The beneficial effects of this invention are: 1. This invention employs a dual-layer closed-loop architecture combining MPC (Model Predictive Control) and high-frequency PID (High-Frequency PID). The dual-loop combination achieves high-precision and robust attitude control of the actuator. Simultaneously, a quantifiable evaluation term for comfort is incorporated into the MPC cost function, transforming comfort from a subjective, fuzzy evaluation into a calculable and optimizable objective goal. During trajectory tracking, the system actively avoids stress concentration, ensuring passenger safety and comfort. Specifically, the MPC controller, based on the system's state-space model, pre-compensates for system inertia and delay, and continuously solves for the optimal future control sequence. The high-frequency PID corrects motor current and speed in real-time with millisecond-level cycles. This dual-loop collaboration ensures that the actuator can still converge to the target pose with high precision under load fluctuations and external disturbances.

[0017] 2. This invention incorporates body pressure heatmaps into a sensing system. It collects human-interface pressure distribution data using arrayed pressure sensors on the seat cushion and backrest. An embedded CNN-LSTM autoencoder, integrating spatiotemporal feature fusion, jointly models the temporal and spatial distribution of body pressure, predicting future pressure distribution trends. The predicted peak body pressure and pressure gradient are then used as quantitative evaluation items for comfort.

[0018] 3. This invention utilizes high-frequency sampling by an angle sensor to feed back the actual pose to the controller, forming a closed-loop correction with the reference trajectory. This compensates for actuator nonlinearity and sensor noise deviation. In each sampling cycle, the MPC (Multi-Process Control) recalculates the optimal control sequence for the future finite time domain based on the current measured state. The system continuously corrects operational deviations and adaptively adjusts the control input according to different user stress response characteristics, proactively optimizing ride comfort while tracking the path. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein: Figure 1 This is a schematic diagram of the overall process logic of an intelligent wheelchair bed posture control method proposed in this invention; Figure 2 This is a diagram of the CNN-LSTM autoencoder network architecture based on spatiotemporal feature fusion in the intelligent wheelchair bed posture control method proposed in this invention. Detailed Implementation

[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0021] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0022] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0023] Example 1

[0024] Reference Figure 1This invention, as an embodiment, is applied to the safety monitoring scenario of elderly people in a virtual nursing home. The wheelchair bed is a core terminal for ensuring the safety of the elderly and for remote monitoring. However, existing devices often rely on preset trajectories or simple logic for posture changes, lacking real-time perception of changes in posture and body pressure distribution, and failing to predict and optimize the posture for the next moment. This leads to mechanical shocks and abrupt movements during bed-chair transitions, causing transient discomfort and even safety hazards. To address this, a method for intelligent wheelchair bed posture control in virtual nursing homes is proposed. It employs a dual-layer closed-loop architecture of Model Predictive Control (MPC) and High-Frequency PID: MPC pre-compensates inertia and delay based on a state-space model and performs rolling optimization; the high-frequency PID corrects motor current and speed in real time with millisecond-level cycles. The dual-loop collaboration ensures high-precision convergence to the target posture even under load fluctuations and disturbances. This solution eliminates shocks and provides compliant movements, offering safe and comfortable adaptive posture control for elderly people in virtual nursing homes, meeting the requirements of remote monitoring for proactive, compliant, and precise execution of intelligent nursing equipment.

[0025] Specifically, the following steps are included: Step 1: Target Position Determination and Smooth Trajectory Planning; Specifically, the system first determines the endpoint target pose for this attitude adjustment based on the target posture parameters input by the user, i.e., the target tilt angle of the backrest relative to the seat surface. And the target folding angle of the thigh and calf plates relative to the seat surface. and .

[0026] To address the technical problems of poor comfort in intermediate postures, instantaneous activation of local muscles, and sudden pressure changes that are easily caused by traditional linear interpolation or constant velocity control, this invention introduces an optimal path search strategy. The implementation method and execution flow of this strategy are as follows: S101: Quantification of discomfort costs and construction of continuous potential fields; The root cause of instantaneous activation of local muscles and sudden changes in pressure lies in the adverse stimulation of human biomechanics by certain specific bed-chair angle combinations. In order to avoid these angles before physical execution, the system first uses a support vector regression algorithm to train a nonlinear surrogate model (such as the existing SVR surrogate model) that converts posture parameters into discomfort scores.

[0027] The system discretizes the entire adjustable angle space of the wheelchair according to a set step size, such as the backrest angle: Thigh angle distance from walking length The grid is discretized (those skilled in the art can adjust the above step size parameters appropriately according to the adjustment accuracy of different wheelchair models or the real-time requirements of different application scenarios); the continuous two-dimensional angle space (backrest angle) is then processed. × Thigh angle The wheelchair is divided into a finite number of discrete grid points, each representing a specific combination of wheelchair postures. This transforms the shortest path search problem in continuous space into a shortest path search problem on a discrete weighted graph, making it solvable by Dijkstra's algorithm. All the grid points generated after discretization become the vertices of the weighted graph in the subsequent Dijkstra optimal path search. Each vertex corresponds to a specific wheelchair posture, and its discomfort cost is predicted and assigned in batches by the SVR surrogate model, thus forming a weighted potential field map covering the global angle space.

[0028] The discomfort cost of all grid points is predicted using the SVR surrogate model, which is the discomfort score, thereby generating an discomfort potential field covering the global space.

[0029] S102: Active obstacle avoidance path optimization based on Dijkstra's algorithm; The system regards each grid point in the discrete potential field as a vertex of a weighted graph, uses an 8-neighborhood connection method to define the motion direction between vertices, and uses the discomfort cost of adjacent vertices as the weight of the edge, and establishes a network graph in combination with the physical constraints of single-step motor operation.

[0030] Dijkstra's algorithm is used to perform a global search on this network graph. Since the algorithm aims to minimize cumulative cost, the planned discrete path node sequence is obtained. It will automatically bypass high-load areas in the potential field, that is, actively avoid intermediate postures that are prone to causing discomfort, and find an optimal posture transformation path that is continuously within a high comfort zone.

[0031] S103: Trajectory Smoothing and Interpolation Processing; Because the discrete grid polyline output by Dijkstra's algorithm can cause discontinuous angular velocities and mechanical jerking when directly driven, the planner combines the maximum angular velocity and angular acceleration constraints of the motor and uses a cubic spline interpolation algorithm to fit the discrete optimal path, ultimately generating a time reference trajectory sequence that satisfies mechanical limits, continuous first derivative (angular velocity), and smoothness. .

[0032] In summary, the smooth optimal reference path planned in this step will serve as the penalty benchmark for the first term (trajectory tracking error) in the Model Predictive Control (MPC) cost function. This ensures that the MPC system will not deviate excessively from this pre-planned "obstacle avoidance backbone" during subsequent real-time rolling re-optimization (actively seeking a comfortable posture). While maintaining overall trajectory change time efficiency, it eliminates the transient force abrupt changes and rapid muscle fatigue caused by traditional open-loop control at the source. Furthermore, the future angle predictions contained in this path will be simultaneously input into the CNN-LSTM network to assist in the accurate prediction of future body pressure distribution.

[0033] Step 2: Discrete state space modeling of the wheelchair bed posture system and motor transmission transfer function; To achieve rolling prediction of future states in Model Predictive Control (MPC) and parameter tuning of the lower-level high-frequency PID controller in subsequent steps, this step requires establishing dynamic mathematical models for each posture adjustment channel (lower leg plate, upper thigh plate, and backrest) of the wheelchair bed. Due to the weak mechanical coupling between the channels, this system identifies their discrete state-space models and motor transfer functions separately. The specific implementation process is as follows: S201: Construct a discrete state-space model for MPC prediction; define the system in... State vector at time step With control input vector as follows: ; ; In the formula, These represent the actual feedback angles of the lower leg plate, upper thigh plate, and back plate at time k. This corresponds to the angular velocity; Let $\mathbf$ and $\mathbf$ be the expected angle increments (i.e., control variables) issued to the motors of the lower leg plate, upper leg plate, and back plate within control cycle $k$. The system conducts identification experiments by applying step control increments at different initial angles and recording the actual angle response curves. Using identification tools, the following discrete state-space equations of the system with a good fit of better than 90% are obtained: ; matrix Meanwhile, the transfer functions of each motor were identified through angle acquisition experiments, which were then used to design the lower-level PID parameters.

[0034] S202: Construct a motor transfer function model for PID control. For the underlying actuators, the system applies drive signals with different PWM duty cycles (e.g., 30%, 50%, 80%) to each channel motor, and collects motor angle change data at high-frequency intervals (e.g., every 20ms). Based on the collected response data, the first-order or second-order transfer function models of the backrest, thigh board, and calf board motors are identified. This provides an accurate controlled object model benchmark for the parameter tuning of the PID controller in subsequent step eight. Based on the obtained transfer functions, the Ziegler-Nichols method is used to calculate the specific PID parameter combinations for each motor under both lifting and lowering conditions, thereby ensuring that the underlying actuators can achieve fast, overshoot-free angle tracking response under different load directions.

[0035] In summary, the discrete state-space equations obtained through step two ( This is the core "prediction engine" of the MPC controller. MPC uses this equation, based on the measured state at the current moment... Recursively predict the attitude state sequence over a finite time domain (N steps). This allows for the prediction and solution of the optimal control increment sequence that minimizes the cost function in step seven. The identified motor transfer functions for each channel are the parameters of the underlying PID controller. The design and Ziegler-Nichols tuning provide a direct benchmark for the controlled object model. This enables the controller designed in step eight to accurately eliminate steady-state errors caused by mechanical hysteresis and static friction within an extremely short control cycle, ensuring that the actuator closely tracks the incremental commands issued by the MPC.

[0036] Step 3: Construct a multi-objective dynamic optimization model predictive control (MPC) cost function; The core of Model Predictive Control (MPC) lies in the fact that at each sampling time k, the current system state can be used as a basis for prediction. The system predicts the system output over a finite time domain (N steps) and solves for the optimal control action using a rolling algorithm. To achieve the dual goals of "accurately tracking the optimal path" and "actively improving comfort," the system uses cubic spline interpolation to generate a discrete-time reference trajectory sequence that matches the predicted time domain, based on the optimal path output by Dijkstra's algorithm in step one. Based on this, the present invention constructs a multi-objective MPC cost function J that includes a ternary weighted cost term: ; The summation terms of the cost function J are as follows: 1. and (State tracking cost): The actual angle vectors of each plate at the future predicted time are obtained by recursion using the discrete state-space equations in step two, where , , These are the actual angles of the lower leg plate, leg plate, and back plate at the future predicted time, obtained by recursion of the discrete state-space equation in step two. This is the corresponding reference planning angle vector. The quadratic form term... Used to calculate the tracking error between the actual path and the optimal path. It constrains the motor to closely follow the main reference trajectory planned in step one, which actively avoids the "high-pressure / high muscle fatigue load area," during dynamic adjustment. Q is the weight matrix for the adjustment angle tracking error and the angular velocity error.

[0037] 2. (Control smoothing cost term): This is the sequence of control increments (i.e., the angle increments of each motor) solved for the system. This quadratic form term... The aim is to punish large abrupt changes in the control increment. By constraining the size of the control variables through the weight matrix R, the motor output is ensured to be smooth and continuous, fundamentally preventing severe physical jerks during the control process.

[0038] 3. (Cost of active comfort optimization): This is the overall objective comfort score estimated by the system based on current and future predicted postures (depending on the prediction results of the random forest comfort assessment model and the CNN-LSTM model in subsequent steps four and five), with a score range of 0 to 100. This nonlinear term... This is the core difference between this invention and traditional MPC. It directly quantifies comfort as a calculable cost, encouraging and driving the optimizer to actively seek and approximate the optimal posture combination with a full comfort score (100 points) within a finite future time domain. w is the comfort deviation penalty coefficient.

[0039] Furthermore, the steps of this invention also include: obtaining weight parameters based on Bayesian optimization. In traditional MPC design, the weight matrices Q, R, and penalty coefficient w often rely on trial and error based on human experience, which is prone to subjective bias and local suboptimal solutions. To address this issue, this invention employs a Bayesian optimization method to globally adaptively optimize and tune the aforementioned parameters. This mechanism successfully locks in the Pareto optimal solution among the three game-like dimensions of "attitude control execution efficiency," "motion smoothness," and "extreme comfort," ensuring the scientific nature and strong robustness of the control system.

[0040] Step 4: Angle sensors deployed in the joints provide real-time feedback on the actual angles of each panel. Simultaneously, pressure sensors embedded in the seat cushion and backrest extract key pressure characteristic parameters, including peak body pressure, average pressure, contact area, and pressure gradient. These characteristic parameters are input into a random forest-based comfort evaluation model (this model has been detailed in Chinese Patent Publication No. CN121786781A, a method for evaluating the static comfort of an intelligent wheelchair based on random forests, and is existing technology, so it will not be elaborated here). The model outputs a quantitative comfort score for the current moment, which serves as the state feedback quantity to update the cost function weights for subsequent control cycles.

[0041] Step 5: Predict future pressure distribution based on spatiotemporal feature fusion network; During the rolling optimization process of Model Predictive Control (MPC), the system needs to accurately evaluate the overall comfort score of each posture in the future prediction time domain. However, it is objectively impossible to obtain the actual physical sensor pressure distribution data corresponding to postures that have not yet occurred. To overcome this physical constraint, this invention constructs a CNN-LSTM autoencoder network based on spatiotemporal feature fusion, aiming to accurately predict the human contact pressure distribution map at the next moment using historical multi-source data sequences. The specific architecture and data feedforward processing flow of this network are as follows: S501: Construction of the input layer of the spatiotemporal feature fusion network; Reference Figure 2 To enable the model to keenly perceive the dynamic pressure transfer trend during changes in human posture, the system input is set to five consecutive historical pressure distribution heatmaps in the time dimension (i.e., ... to The specifications of the input matrix for each frame are exactly the same as those of the real sensor array. A single-channel matrix. Simultaneously, the system synchronously inputs the wheelchair posture angle sequence corresponding to each historical frame.

[0042] S502: Spatial topological feature extraction based on CNN encoder; The model front end uses a convolutional neural network (CNN) as the encoder, and performs independent spatial topological feature extraction on the stress maps of the above 5 consecutive frames through temporally distributed operations. The input matrix is ​​sequentially processed through two steps with a stride of 2 and padding of 1. Convolutional layer. The first convolutional layer expands the number of channels to 16, and the feature map size is downsampled. The second convolutional layer further reduces its dimensionality to... Subsequently, the feature map is flattened into a one-dimensional vector of length 2^112, and then compressed to a low-dimensional feature representation of 60 dimensions by a fully connected layer (FC). S503: Time-dependent evolution modeling based on LSTM; The extracted spatial feature vector sequences at five independent time points are sequentially input into a Long Short-Term Memory (LSTM) network. The LSTM layer configuration of this invention contains 60 hidden units. After five time steps of sequential deep learning, the LSTM can effectively capture the dynamic evolutionary dependence of the pressure distribution in the lower back and hips over time, and output the final hidden state vector h(t) as the core vector representing the spatiotemporal context of the future prediction step.

[0043] S504: Temporal Dependent Evolutionary Modeling Based on LSTM Reconstructed from a Pressure Heatmap Using a CNN Decoder; the decoder is responsible for accurately mapping the low-dimensional spatiotemporal representation h(t) back to the original physical space dimension to generate the predicted image. First, the 60-dimensional features are upsized to 2^112 using a fully connected layer and then reshaped. The high-dimensional tensor is then used for upsampling and reconstruction through two transposed convolutional layers. The first transposed convolutional layer outputs... The first layer features the characteristic map, and the second layer accurately restores it to a single-channel predicted pressure matrix P(t+1) consistent with the dimensions of the real sensor.

[0044] S505: Feature Deep Concatenation and Future Comfort Calculation; After the network generates a high-resolution predicted pressure matrix P(t+1), the system extracts 28 key pressure features (such as peak body pressure, contact area, pressure gradient, etc.) from this matrix. These predicted pressure features are fed into the random forest evaluation model trained in step 4 to accurately calculate the overall comfort prediction score under future postures. This score will be immediately fed back into the MPC cost function in step three, giving the MPC system the ability to proactively optimize high-comfort postures with feedforward control.

[0045] Step Six: Setting Physical Constraints and Control Increment Boundary Conditions for the Actuator; When model predictive control (MPC) is used to solve for the optimal sequence, in order to ensure that the motion commands generated by the algorithm strictly conform to the actual physical execution capability of the wheelchair motor, and to prevent excessive single-step adjustment from causing harsh physical impact and transient discomfort to the human body, the system must impose explicit amplitude constraints on the control variable (angle increment). Specifically, for each step of the control increment command within the prediction time domain, the system sets the following linear inequality constraint: ; in, This indicates the control amount for a single adjustment of the backplate. This indicates the control amount for a single adjustment of the legboard. This indicates the control amount of a single adjustment of the calf plate. , representing the control quantity at each step within the finite prediction time domain in the future.

[0046] By setting this physical constraint, mechanical jerking and impact sensation can be prevented. Specifically, the mechanical impact sensation of the wheelchair during bed-chair posture changes directly originates from sudden acceleration or sharp turns of the actuator. Through the aforementioned explicit amplitude constraint, the system enforces that the single-step extreme value of the backrest motor does not exceed... The single-step operating extreme value of the thigh plate motor does not exceed This design ensures a smooth adjustment range from the source, effectively suppressing sudden changes in the dynamic pressure rate and preventing instantaneous compression of the user's local soft tissues.

[0047] Furthermore, the aforementioned magnitude constraint inequalities, along with the cost function constructed in step three, will be used as core input parameters in the subsequent Sequence Quadratic Programming (SQP) solver in step seven. These inequalities limit the search space of the optimization variables, ensuring not only the physical executability of the solved control sequence but also significantly improving the efficiency of the algorithm's rolling optimization.

[0048] Step 7: Rolling time-domain solution and optimal control sequence output based on Sequence Quadratic Programming (SQP); After obtaining the current system state feedback and establishing the cost function in the prediction time domain, Model Predictive Control (MPC) enters the core optimization and solution phase. This step integrates all the aforementioned models, constraints, and prediction data to calculate the optimal motor action command for the current control cycle in real time. The specific execution flow is as follows: S701: Reconstructing the functional expression of the optimization problem; converting the discrete state-space equations established in step two ( Substituting these values ​​into the multi-objective cost function J constructed in step three, and utilizing the model's recursive relationship, the system state and output for the next N steps are expressed as a whole as the current measured initial state. With future control sequences The function.

[0049] S702: Linearization of nonlinear costs and solution of convex quadratic programming (QP); In the reconstructed cost function J, the first term (angle tracking error) and the second term (control increment penalty) belong to the standard quadratic form with respect to the control quantity U, but the third term is the comfort score. It is highly nonlinear.

[0050] To meet the real-time solution requirements of the system, this invention employs the Sequential Quadratic Programming (SQP) method to handle this problem: at each sampling time, the nonlinear comfort function is... A first-order Taylor expansion is performed near the current control sequence operating point, thereby locally approximating the originally complex nonlinear optimization problem as a standard convex quadratic programming (QP) problem with linear constraints (in conjunction with step six), and then iteratively solving the problem using a QP solver until convergence.

[0051] S703: Rolling control mechanism and command output; using the solver to obtain the optimal control increment sequence for the next N steps. Subsequently, following the rolling time-domain control principle, the system does not execute the entire sequence, but only extracts the first control variable from the sequence. As the actual control command for the current period, it is directly sent to the underlying PID controller (step eight) for execution.

[0052] Step 8: High-frequency fine closed-loop control of the actuator based on PID algorithm; Model predictive control (MPC) requires deep CNN-LSTM inference and multi-objective SQP rolling optimization, resulting in a relatively long control optimization cycle (set at 200ms). This makes it difficult to directly and promptly address transient high-frequency disturbances caused by mechanical backlash, frictional resistance, and human center of gravity shift during the underlying motor operation. Therefore, this invention constructs a "dual-timescale" control architecture, introducing a high-frequency PID closed-loop controller with an extremely short control cycle (40ms) at the execution level. The specific execution scheme is as follows: S801: Definition of underlying control law and error model; the system is equipped with independent standard PID controllers for the backrest motor, thigh board motor, and calf board motor. The control law calculation formula is as follows: ; In the formula, e(t) is the control deviation, which is defined as: the expected control quantity issued by MPC in the current control cycle (i.e., the optimal angle increment command obtained by rolling solution in step seven). The difference between the actual execution angle increment fed back in real time by the underlying motor encoder within the cycle, For controller output, To control deviation, , , These are all parameters for a PID controller.

[0053] S802: Parameter Acquisition; Considering the significant physical differences in the human body and mechanical gravity load characteristics that the motor overcomes during the "lifting" and "lowering" movements of the wheelchair bed, this invention abandons the design approach of a single static parameter. Based on the specific motor transfer function identified in step two, the system uses the Ziegler-Nichols method to tune exclusive PID parameter combinations for each part of the motor under the two different dynamic conditions of "lifting" and "lowering". For example, in this embodiment, the thigh plate uses a low-proportion, low-differential parameter when lifting to gently overcome gravity, while using a high-proportion, high-differential parameter when descending to quickly suppress the tendency to accelerate with gravity.

[0054] This embodiment also provides an intelligent wheelchair bed posture control system, which includes: a planner module, an MPC controller module, a spatiotemporal feature fusion network module, and a PID controller module.

[0055] The planner module combines the motor's maximum angular velocity and angular acceleration constraints, uses cubic spline interpolation to fit the discrete optimal path, and generates a continuous and smooth time reference trajectory sequence, which is used as the penalty benchmark for the first term in the MPC cost function. The MPC controller module combines the discrete state-space equation and the MPC cost function, and solves the optimal control sequence output based on the rolling time domain of sequential quadratic programming. The spatiotemporal feature fusion network module generates a high-resolution predicted pressure matrix, extracts pressure features based on this predicted pressure matrix, and feeds them into a trained random forest evaluation model to calculate the overall comfort prediction score under future postures. This score serves as the active comfort optimization cost term in the MPC cost function. The PID controller module uses the difference between the expected control quantity of the optimal control sequence and the actual execution angle increment fed back by the underlying motor encoder in real time within the cycle as the control error input to the PID control law for wheelchair posture adjustment.

[0056] It should be noted that the above modules can be functional modules or program modules, and can be implemented through software or hardware. For modules implemented through hardware, the above modules can reside in the same processor; or the above modules can be located in different processors in any combination.

[0057] The information interaction and execution process between the above modules are based on the same concept as the method embodiment of this application. It is a system corresponding to the above-mentioned intelligent wheelchair bed posture control method. All implementation methods in the above method embodiment are applicable to the system embodiment. For its specific functions and technical effects, please refer to the method embodiment section, which will not be repeated here.

[0058] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements a method for assessing the safety risks of highway construction as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0059] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for adjusting the posture of an intelligent wheelchair bed, characterized in that, Includes the following steps: Constructing discrete state-space equations for MPC prediction: ; in, for State vector at time step: ; To control the input vector: ; In the formula, These represent the actual feedback angles of the lower leg plate, upper thigh plate, and back plate at time k. This corresponds to the angular velocity; These represent the expected angle increments sent to the lower leg plate, upper leg plate, and back plate motors within control cycle k, respectively. Constructing the multi-objective MPC cost function J: ； In the formula, The actual angle vectors of the lower leg plate, leg plate, and back plate; This is the corresponding reference planning angle vector; To solve for the control increment sequence, The overall objective comfort score is estimated based on the current and future predicted postures. Q and R are the corresponding weight matrices, w is the comfort deviation penalty coefficient, and N is the prediction step size. Combining discrete state-space equations, the cost function is solved in the rolling time domain based on sequential quadratic programming. The minimized optimal control increment sequence; Each of the backrest motor, thigh board motor, and calf board motor is equipped with an independent standard PID controller. The difference between the actual execution angle increment and the real-time feedback from the underlying motor encoder within the cycle is determined by the optimal control increment sequence as the control deviation, which is then input to the corresponding PID controller to achieve posture control.

2. The method for adjusting the posture of an intelligent wheelchair bed according to claim 1, characterized in that: The The acquisition method is as follows: the actual angle of each plate is fed back in real time by angle sensors deployed in the joints; and key pressure characteristic parameters, including peak body pressure, average pressure, contact area and pressure gradient, are extracted by pressure sensors embedded in the seat cushion and backrest. The feature parameters are input into a random forest-based comfort evaluation model, which outputs a quantitative score of comfort at the current moment in real time.

3. The method for adjusting the posture of an intelligent wheelchair bed according to claim 2, characterized in that: The method also includes: constructing a CNN-LSTM autoencoder network based on spatiotemporal feature fusion, and using this CNN-LSTM autoencoder network to predict the human contact pressure distribution map at the next time step; specifically including: A convolutional neural network was used as the encoder, and spatial topological features were extracted independently from the stress maps of five consecutive frames through temporally distributed operations. The input matrix was sequentially processed through two steps with a stride of 2 and padding of 1. Convolutional layer; the first convolutional layer expands the number of channels to 16, and the feature map size is downsampled to [value missing]. The second convolutional layer further reduces its dimensionality to... The feature map is flattened into a one-dimensional vector of length 2^112, and then compressed to a low-dimensional feature representation of 60 dimensions by a fully connected layer. The extracted spatial feature vector sequences at 5 independent time points are sequentially input into the Long Short-Term Memory network to output the final hidden state vector h(t). The decoder is responsible for accurately mapping the low-dimensional spatiotemporal representation h(t) back to the original physical space dimension, generating the predicted image: the 60-dimensional features are upscaled to 2^112 through a fully connected layer and reshaped into... A high-dimensional tensor; upsampling and reconstruction are performed through two transposed convolutional layers; the first transposed convolutional layer outputs... The feature map is then precisely restored to a single-channel predicted pressure matrix P(t+1) consistent with the dimensions of the real sensor in the second layer. After the network generates a high-resolution predicted stress matrix P(t+1), key stress features are extracted from this predicted matrix. Then, a random forest evaluation model is used to calculate the overall comfort prediction score for future postures. .

4. The method for adjusting the posture of an intelligent wheelchair bed according to claim 1, characterized in that: The method combines discrete state-space equations and uses rolling time-domain solutions based on sequential quadratic programming to obtain the cost function. When minimizing the optimal control increment sequence, a linear inequality constraint is set for the control increment command at each step in the prediction time domain: ； in, This indicates the control amount for a single adjustment of the backplate. This indicates the control amount for a single adjustment of the legboard. This indicates the control amount of a single adjustment of the calf plate. , representing the control quantity at each step within the finite prediction time domain in the future.

5. The method for adjusting the posture of an intelligent wheelchair bed according to claim 1, characterized in that: The control law of the PID controller is designed as follows: ； In the formula, e(t) represents the control deviation. For controller output, To control deviation, , , These are all parameters for a PID controller.

6. The method for adjusting the posture of an intelligent wheelchair bed according to claim 1, characterized in that: The method also includes: Based on the target posture parameters input by the user, determine the target tilt angle of the backrest relative to the seat surface. And the target folding angle of the thigh and calf plates relative to the seat surface. and ; The entire adjustable angle space of the wheelchair is discretized into a grid according to the set step length, and the continuous two-dimensional angle space is divided into a finite number of discrete grid points. The discomfort cost of all grid points is predicted using the SVR surrogate model, and an discomfort potential field covering the global space is generated. Each discrete grid point is regarded as a vertex of a weighted graph. The motion direction between vertices is defined by an 8-neighborhood connection method. The discomfort score of adjacent vertices is used as the weight of the edge. The network graph is established in combination with the physical constraints of the motor's single-step operation. Dijkstra's algorithm is used to perform a global search in the network graph, and the resulting discrete path node sequence is planned. It will automatically avoid high-load areas in the potential field; The planner combines the motor's maximum angular velocity and angular acceleration constraints, and uses a cubic spline interpolation algorithm to fit the discrete optimal path to generate a reference trajectory sequence: This is used as the penalty benchmark for the first term in the MPC cost function.

7. The method for adjusting the posture of an intelligent wheelchair bed according to claim 1, characterized in that: The method also includes: constructing a motor transfer function model for PID control, specifically: By applying drive signals with different PWM duty cycles to each channel motor, the angle change data of the motor is collected at a high frequency cycle. Based on the collected angle change data, the first-order or second-order transfer function models of the backrest, thigh board, and calf board motors are identified. Based on the obtained transfer function models, the Ziegler-Nichols method is used to calculate the exclusive PID parameter combinations of each motor under the two working conditions of lifting and lowering.

8. The method for adjusting the posture of an intelligent wheelchair bed according to claim 1, characterized in that: The weight matrices Q and R, as well as the penalty coefficient w, are globally adaptively tuned using a Bayesian optimization method.

9. An intelligent wheelchair bed posture control system, applied to the intelligent wheelchair bed posture control method according to any one of claims 1-8, characterized in that: The system includes: The planner module combines the maximum angular velocity and angular acceleration constraints of the motor, uses a cubic spline interpolation algorithm to fit the discrete optimal path to generate a continuous and smooth time reference trajectory sequence, and uses it as the penalty benchmark for the first term in the MPC cost function. The MPC controller module is used to combine discrete state-space equations with the MPC cost function to solve for the optimal control sequence output in the rolling time domain based on sequential quadratic programming. The spatiotemporal feature fusion network module is used to generate a high-resolution predicted stress matrix. Based on the predicted stress matrix, stress features are extracted and fed into the trained random forest evaluation model to calculate the overall comfort prediction score under future postures. This score is used as the active comfort optimization cost term in the MPC cost function. The system also includes a PID controller module, which uses the difference between the expected control quantity of the optimal control sequence and the actual execution angle increment fed back in real time by the underlying motor encoder within the cycle as the control error input to the PID control law to perform wheelchair bed posture adjustment.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the intelligent wheelchair bed posture control method according to any one of claims 1 to 8.

Images